Emad Jaha Thesis Draft v12 - COnnecting REpositories · Title: Microsoft Word - Emad Jaha_Thesis...
Transcript of Emad Jaha Thesis Draft v12 - COnnecting REpositories · Title: Microsoft Word - Emad Jaha_Thesis...
ii
DEDICATION
I would like to dedicate my thesis to
My Parents
My Wife
and my two children
iii
ACKNOWLEDGMENT
All my thanks are due to Almighty Allah (Subhanahu Wa Taalaa) who gave me courage
and patience to carry out this work.
I would like to thank Dr. Lahouari Ghouti for his precious time and continuous advice,
support and encouragement throughout this research. Also, my thesis committee
members, Dr. Nasir Al-Darwish and Dr. Adel Fadhl Ahmed, provided very helpful
advice, comments and feedback to improve the content of this research.
Furthermore, I would like to thank all faculty members, colleagues and friends who
shared with me the path of learning and knowledge throughout my graduate studies.
I wish to express my gratitude to my family and my wife’s family members for their
encouragement and being patient with me.
Special thanks are to King Abdul-Aziz University (KAU) for their generous support and
scholarship to complete my Master degree and to King Fahd University of Petroleum and
Minerals (KFUPM) for their support of this research.
iv
TABLE OF CONTENT Page
DEDICATION ................................................................................................................................. ii
ACKNOWLEDGMENT ................................................................................................................. iii
TABLE OF CONTENT .................................................................................................................. iv
LIST OF TABLES .......................................................................................................................... vi
LIST OF FIGURES ........................................................................................................................ vii
THESIS ABSTRACT ...................................................................................................................... ix
x ....................................................................................................................................... ملخص الرسالة
CHAPTER 1 ........................................................................................................................................ 1
1. INTRODUCTION .................................................................................................................... 1
1.1. Biometrics ......................................................................................................................... 3
1.2. Face Recognition .............................................................................................................. 4
1.2.1. Face Recognition Processing .................................................................................... 6
1.3. Problem Statement ............................................................................................................ 8
1.4. Motivation ........................................................................................................................ 8
1.5. Thesis Contributions ......................................................................................................... 9
1.6. Thesis Outline ................................................................................................................. 10
CHAPTER 2 ...................................................................................................................................... 11
2. LITERATURE REVIEW ....................................................................................................... 11
2.1. Grayscale-based Face Recognition ................................................................................. 11
2.1.1. Grayscale PCA-based Approach ............................................................................ 11
2.1.2. Grayscale Gabor Filters Approach ......................................................................... 12
2.1.3. Grayscale Wavelet-based Approach ....................................................................... 13
2.2. Color-based Face Recognition ........................................................................................ 13
2.2.1. Color PCA-based Approach ................................................................................... 15
2.2.2. Color Quaternion-based Approach ......................................................................... 15
CHAPTER 3 ...................................................................................................................................... 19
3. QUATERNION (HYPERCOMPLEX) REPRESENTATION .............................................. 19
3.1. Quaternion Numbers....................................................................................................... 19
3.2. Properties of Quaternion Numbers ................................................................................. 20
3.3. Quaternion Representation of Color Face Images .......................................................... 22
v
3.4. Quaternion Fourier Transform (QFT) ............................................................................ 23
3.5. Quaternion Gabor Filters ................................................................................................ 25
CHAPTER 4 ...................................................................................................................................... 26
4. PRINCIPAL COMPONENT ANALYSIS (PCA).................................................................. 26
4.1. Eigenfaces ....................................................................................................................... 27
4.2. Solving PCA Using SVD ............................................................................................... 28
4.3. Quaternion Principal Component Analysis (Q-PCA) ..................................................... 28
4.3.1. Quaternion Singular Value Decomposition (Q-SVD) ............................................ 29
4.3.2. QSVD-Based Color Image Compression ............................................................... 32
4.3.3. Proposed Q-PCA Algorithm ................................................................................... 34
CHAPTER 5 ...................................................................................................................................... 36
5. METHODOLOGY AND EXPERIMENT ............................................................................. 36
5.1. Research Methodology ................................................................................................... 36
5.2. Color Face Image Database ............................................................................................ 36
5.2.1. Georgia Tech Face Database (GTDB) .................................................................... 37
5.2.2. Color FERET Database .......................................................................................... 37
5.3. Performance Evaluation ................................................................................................. 39
5.3.1. Standard Testing Subsets ........................................................................................ 40
5.3.2. Test Design ............................................................................................................. 42
5.3.3. Standard Performance Measures ............................................................................ 43
5.3.4. Identification Performance against Probe Categories ............................................. 49
5.3.5. Performance Comparison Using CMC in Different Resolutions ........................... 57
5.3.6. Variation in Identification Performance ................................................................. 72
5.3.7. GD versus ID .......................................................................................................... 79
5.3.8. FAR versus FRR errors .......................................................................................... 84
5.3.9. ROC Curves ............................................................................................................ 89
CHAPTER 6 ...................................................................................................................................... 94
6. CONCLUSION AND FUTURE WORK ............................................................................... 94
6.1. Summary and Findings ................................................................................................... 94
6.2. Future Work .................................................................................................................... 97
REFERENCES ............................................................................................................................... 99
Vita ............................................................................................................................................... 104
vi
LIST OF TABLES
Page
Table 5.1: Summary of FERET color face database....................................................................... 39
Table 5.2: Summary of PCA-based algorithms used in this thesis................................................. 40
Table 5.3: Gallery and four probe sets............................................................................................ 40
Table 5.4: Resolution of gallery and probe sets.............................................................................. 40
Table 5.5: Subsets used to investigate the variation in performance.............................................. 42
Table 5.6: Summary of used gallery and probe categories............................................................. 49
Table 5.7: Match scores of the top-rank for each PCA-based algorithm........................................ 50
Table 5.8: Average score of the total match scores of rank 50 for each PCA-based algorithm..... 50
Table 5.9: Average score of the full rank for each PCA-based algorithm in two resolutions........ 57
Table 5.10: Variation in identification performance using five different galleries from the FB
probes (top match).......................................................................................................................... 72
Table 5.11: Variation in identification performance using five different galleries from the FB
probes (average of rank scores)...................................................................................................... 73
Table 5.12: Variation in identification performance using five different galleries from the
duplicate I probes (top match)........................................................................................................ 76
Table 5.13: Variation in identification performance using five different galleries from the
duplicate I probes (average of rank scores).................................................................................... 76
Table 5.14: Decidability index (d) for each algorithm.................................................................... 79
Table 5.15: The EER of each algorithm......................................................................................... 84
vii
LIST OF FIGURES
Page
Figure 1.1: Main processing steps in pattern recognition systems. .................................................. 2 Figure 1.2: Biometric categories. ..................................................................................................... 3 Figure 1.3: Categorization of Face Recognition (FR) methods. ....................................................... 6 Figure 1.4: Face recognition processing flow. ................................................................................. 6 Figure 1.5: Face detection and alignment processes. ....................................................................... 7 Figure 1.6: Four areas of related work and the current research location. ....................................... 9 Figure 3.1: Quaternion representation of a color face image. ........................................................ 22 Figure 4.1: The PCA concept, (Li et al. [14]). ................................................................................ 26 Figure 4.2: In the top gray-eigenfaces and in the bottom color-eigenfaces. ................................... 28 Figure 4.3: Q-SVD-based image compression. (a),(b),(c) and (d) are the reconstructed images with k=8,16,25,144 respectively. Where (d) is the perfect reconstruction of the original image ........... 34 Figure 5.1: One subject face images (gallery and probes). ............................................................. 41 Figure 5.2: Design scheme of the testing procedure....................................................................... 45 Figure 5.3: GD versus ID and FAR versus FRR errors. ................................................................. 47 Figure 5.4: FAR versus FRR curves. .............................................................................................. 48 Figure 5.5: Identification performance using the FB with 192×128 pixels resolution. .................. 53 Figure 5.6: Identification performance using the FB with 384×256 pixels resolution. .................. 53 Figure 5.7: Identification performance using the duplicate I with 192×128 pixels resolution. ...... 54 Figure 5.8: Identification performance using the duplicate I with 384×256 pixels resolution. ...... 54 Figure 5.9: Identification performance using the fc with 192×128 pixels resolution. .................... 55 Figure 5.10: Identification performance using the fc with 384×256 pixels resolution. .................. 55 Figure 5.11: Identification performance using the duplicate II with 192×128 pixels resolution. .. 56 Figure 5.12: Identification performance using the duplicate II with 384×256 pixels resolution. .. 56 Figure 5.13: The CMC curves for PCA using the FB probes. ....................................................... 58 Figure 5.14: The CMC curve for PCA using the duplicate I probes. ............................................. 58 Figure 5.15: The CMC curves for PCA using the fc probes. ......................................................... 59 Figure 5.16: The CMC curves for PCA using the duplicate II probes. .......................................... 59 Figure 5.17: The CMC curves for QPCA using the FB probes. ..................................................... 60 Figure 5.18: The CMC curves for QPCA using the duplicate I probes. ......................................... 60 Figure 5.19: The CMC curves for QPCA using the fc probes. ....................................................... 61 Figure 5.20: The CMC curves for QPCA using the duplicate II probes. ....................................... 61 Figure 5.21: The CMC curves for Red-PCA using the FB probes. ................................................ 62 Figure 5.22: The CMC curves for Red-PCA using the duplicate I probes. .................................... 62 Figure 5.23: The CMC curves for Red-PCA using the fc probes. .................................................. 63 Figure 5.24: The CMC curves for Red-PCA using the duplicate II probes. ................................... 63 Figure 5.25: The CMC curves for Green-PCA using the FB probes. ............................................. 64 Figure 5.26: The CMC curves for Green-PCA using the duplicate I probes. ................................. 64 Figure 5.27: The CMC curves for Green-PCA using the fc probes. .............................................. 65 Figure 5.28: The CMC curves for Green-PCA using the duplicate II probes. ............................... 65
viii
Figure 5.29: The CMC curves for Blue-PCA using the FB probes. ............................................... 66 Figure 5.30: The CMC curves for Blue-PCA using the duplicate I probes. ................................... 66 Figure 5.31: The CMC curves for Blue-PCA using the fc probes. ................................................. 67 Figure 5.32: The CMC curves for Blue-PCA using the duplicate II probes. ................................. 67 Figure 5.33: The CMC curves for Avg RGB-PCA using the FB probes. ...................................... 68 Figure 5.34: The CMC curves for Avg RGB-PCA using the duplicate I probes. .......................... 68 Figure 5.35: The CMC curves for Avg RGB-PCA using the fc probes. ........................................ 69 Figure 5.36: The CMC curves for Avg RGB-PCA using the duplicate II probes. ......................... 69 Figure 5.37: The CMC curves for SCFT-CPCA using the FB probes. .......................................... 70 Figure 5.38: The CMC curves for SCFT-CPCA using the duplicate I probes. .............................. 70 Figure 5.39: The CMC curves for SCFT-CPCA using the fc probes. ............................................ 71 Figure 5.40: The CMC curves for SCFT-CPCA using the duplicate II probes. ............................. 71 Figure 5.41: Identification performance using the FB probes of (Set 1) and rank of 50 ................ 73 Figure 5.42: Identification performance using the FB probes of (Set 2) and rank of 50 ................ 74 Figure 5.43: Identification performance using the FB probes of (Set 3) and rank of 50 ................ 74 Figure 5.44: Identification performance using the FB probes of (Set 4) and rank of 50 ................ 75 Figure 5.45: Identification performance using the FB probes of (Set 5) and rank of 50 ................ 75 Figure 5.46: Identification performance using the duplicate I probes of (Set 1) and rank of 50 .... 77 Figure 5.47: Identification performance using the duplicate I probes of (Set 2) and rank of 50 .... 77 Figure 5.48: Identification performance using the duplicate I probes of (Set 3) and rank of 50 .... 78 Figure 5.49: Identification performance using the duplicate I probes of (Set 4) and rank of 50 .... 78 Figure 5.50: Identification performance using the duplicate I probes of (Set 5) and rank of 50 .... 79 Figure 5.51: GD versus ID for PCA algorithm .............................................................................. 81 Figure 5.52: GD versus ID for QPCA algorithm ............................................................................ 81 Figure 5.53: GD versus ID for Red PCA algorithm ....................................................................... 82 Figure 5.54: GD versus ID for Green PCA algorithm .................................................................... 82 Figure 5.55: GD versus ID for Blue PCA algorithm ...................................................................... 83 Figure 5.56: GD versus ID for Avg RGB PCA algorithm ............................................................. 83 Figure 5.57: GD versus ID for SCFT CPCA algorithm ................................................................. 84 Figure 5.58: FAR versus FRR curve for PCA algorithm ............................................................... 86 Figure 5.59: FAR versus FRR curves for QPCA algorithm ........................................................... 86 Figure 5.60: FAR versus FRR curves for Red PCA algorithm ...................................................... 87 Figure 5.61: FAR versus FRR curves for Green PCA algorithm ................................................... 87 Figure 5.62: FAR versus FRR curves for Blue PCA algorithm ..................................................... 88 Figure 5.63: FAR versus FRR curves for Avg RGB PCA algorithm ............................................. 88 Figure 5.64: FAR versus FRR curves for SCFT CPCA algorithm ................................................. 89 Figure 5.65: PCA and QPCA algorithms ROC curves ................................................................... 90 Figure 5.66: ROC curve for Red PCA algorithm ........................................................................... 91 Figure 5.67: ROC curve for Green PCA algorithm ........................................................................ 91 Figure 5.68: ROC curve for Blue PCA algorithm .......................................................................... 92 Figure 5.69: ROC curve for Avg RGB PCA algorithm.................................................................. 92 Figure 5.70: ROC curve for SCFT CPCA algorithm ..................................................................... 93
ix
THESIS ABSTRACT
Name: Emad Sami Jaha
Title: Color Face Recognition Using Quaternion Principle Component
Analysis (Q-PCA)
Major Field: Information and Computer Science
Date of Degree: November 2010
Recently, biometric systems have attracted the attention of both academic and industrial
communities. Advances in hardware and software technologies have paved the way to
such growing interest. Nowadays, efficient and cost-effective biometric solutions are
continuously emerging. Fingerprint-based biometric systems have pioneered the
commercial applications. Face and iris traits have been proven to be reliable candidates.
Until recently, face recognition research literally followed the research undertaken in the
field of fingerprint recognition which is inherently gray-scale. In this research, efforts are
restricted to the investigation of subspace representations in the color domain. The
concept of principal component analysis (PCA) implemented via singular value
decomposition (SVD) is carried over into the hypercomplex (i.e., quaternionic) to define
quaternionic PCA (Q-PCA) where color faces are compactly represented. Unlike the
existing approaches for handling the color information, the proposed research implicitly
accounts for the correlation that exists between the color components (i.e., red, green and
blue components).
x
ملخص الرسالة
عماد بن سامي بن محمد جاها :االســـــــــــــــم
التعرف على صور الوجوه الملونة باستخدام تحليل المكونات األساسية المرآبة الرباعية :عنوان الرسالة
األبعاد
علوم الحاسب اآللي :التخصـــــــص
هـ1431ذو القعدة :تاريخ التخـرج
إن . األشخاصهوية في اآلونة األخيرة انصرف اهتمام األوساط األآاديمية والصناعية لنظم التعرف الحيوية على
ظهرت و أليام في هذه اوبالتالي . التطور في تقنيات العتاد والبرمجيات قد مهد الطريق لمثل هذا االهتمام المتزايد
الحيوية المعتمدة على التعرفنظم إن. اقتصادية وعالية الكفاءة حلولبشكل مستمر الكثير من نظم التعرف الحيوية آ
وقد ثبت أن سمات الوجه والقزحية مرشح يمكن االعتماد .بصمات األصابع آان لها السبق في التطبيقات التجارية
حتى وقت قريب آانت البحوث في مجال التعرف على صور الوجوه تتبع . عليه في بناء مثل هذه النظم الحيوية
حرفيا مسار البحوث التي أجريت في مجال التعرف على بصمات األصابع والتي تعتمد بطبيعتها على الصور الممثلة
أما في هذا البحث تقتصر الجهود على مساحة جزئية لتمثيل الصور في نطاق . تدرجات اللون الرمادي في نطاق
التي يتم حسابها بواسطة تحليل القيم المفردة ومن ثم تمثيل صور الوجوه المكونات األساسيةمفهوم تحليل إن . األلوان
األساليب الموجودة لمعالجة معلومات األلوان، فإن بخالف. بشكل مدمج باستخدام الصيغة المرآبة الرباعية األبعاد
.)األحمر واألخضر واألزرقأي (البحث المقترح يضع في الحسبان االرتباط القائم بين األلوان المكونة للصورة
1
CHAPTER 1
1. INTRODUCTION
The revolution in the realm of modern technologies has influenced explicitly the people
daily life. Technology has become widely used in many forms and everywhere. In fact, it
has facilitated several needs such as communication, automated processes, and enforced
security. Many of existing security-based systems such as banking systems are using
different forms of technology. In such sensitive systems, a facility for user authentication
and identification is considered an urgent need. Several of the commonly used solutions
are username/password, electronic and smart cards. Although these solutions are effective
and have many advantages, they suffer from serious uncontrollable disadvantages. Several
violations could be easily performed such as using a stolen username/password or
electronic/smart card by unauthorized persons. Another drawback consists in losing or
forgetting the identity mean (i.e., username/password or smart card), which can also
prevent the legal person from accessing the system. Also, other various forms of
violations and problems can occur effortlessly.
Due to such disadvantages, biometric solutions have taken place and started playing a
main role in developing efficient authentication and identification systems which are free
from the above mentioned drawbacks. Biometric-based systems make use of various
unique human traits. The majority of earlier research efforts have focused on fingerprint-
based biometric systems. After that, face and iris traits have become the focus of very
active research. Face recognition (FR) has many advantages over other biometric traits
because it is natural, nonintrusive, and easy to use. Face recognition research has literally
2
followed fingerprint recognition research in methodologies and techniques used; therefore
most of published research has used inherently gray-scale images. However, a number of
promises were carried out from research which has used color images for FR purposes.
The FR problem can be cast into the broader area of pattern recognition.
A simplified description of any automated pattern recognition system is using a computer
with its integrated components (either hardware or software), instead of human
observation, to recognize or identify the different samples by using a previous knowledge
of similar models.
Any pattern recognition system comprises three main steps: preprocessing, feature
extraction, and classification, as shown in Figure 1.1. Preprocessing is a set of prior
operations applied, as preparation, on the input data such as resizing, normalizing,
filtering, localization, and normalization. Features are the individual measurable heuristic
properties of the phenomena being observed. These properties are used within
comparisons between a new unknown sample and already known models to recognize to
which class it belongs. Classifiers are mechanisms to assign each point that represents a
pattern in the space with a class label or membership scores to the defined classes.
Figure 1.1: Main processing steps in pattern recognition systems.
3
1.1. Biometrics
A biometric is defined as “a measurable, physical characteristic or personal behavioral
trait used to recognize the identity, or verify the claimed identity, of an enrollee” [1].
There exist several different biometrics comprise some methods for uniquely recognizing
humans based on one or more essential (innate) physical or behavioral characteristic.
Biometric characteristics can be classified into two main categories listed below and
depicted in Figure 1.2:
Physiological biometrics are related to the shape of the body, e.g. fingerprint, face and
iris recognition, hand and palm geometry, and DNA.
Behavioral biometrics are related to the behavior of a person, e.g. typing rhythm or
Keystroke, gait, signature, and voice.
Figure 1.2: Biometric categories.
Biometric
Physiological
Face
Fingerprint
Hand
Iris
DNA
Behavioral
Keystroke
Gait
Signature
Voice
4
A biometric system can operate in the following two modes [4]:
Verification (or authentication) which is a one-to-one matching or comparison of a
captured biometric with a stored template to verify that the individual is the one he
claimed by comparing a query face image against a template face image.
Identification (or recognition) which is a one-to-many matching or comparison of the
captured biometric against all the template images in a biometric database in attempt to
identify an unknown individual, where a query face is matched to a list of suspects can
be one-to-few matches.
1.2. Face Recognition
A face recognition (FR) system is simply an attempt to emulate the typical human face
recognition task that a human performs routinely, effortlessly, and frequently along his
life. Thus face recognition defined as the ability of a computer to receiving and
interpreting of face image input. Such face recognition system is supposed to identify
faces presented in images and videos automatically. It can operate in either a single or
dual mode. Like other biometric systems, an FR system is capable of [1]:
1) Capturing a face sample from an end user.
2) Extracting biometric data from that sample.
3) Comparing the biometric data with that contained in one or more reference
templates.
4) Deciding how well they match.
5
5) Indicating whether or not an identification or verification of identity has been
achieved.
Face recognition has undergone approximately fifteen years of intensive development in
both the academic and commercial arenas [1]. Despite of there exist several different
approaches in the academic literature, they can be divided into four major classes of
algorithms:
1) Eigenface Systems based on eigenfaces and eigenvectors.
2) Local Feature Analysis System based on the analysis of local features.
3) Neural Network methods based on machine learning techniques.
4) Gabor filter methods based on Gabor and wavelets analogies.
Definitely, each class has its own advantages and disadvantages [1]. The diagram in,
shown in Figure 1.3, demonstrates classification of FR methods.
Although, FR is a very popular and promising research topic, the task is also tending to be
a difficult one due to the existence of unconstrained tasks such as viewpoint, illumination,
expression, occlusion, and accessories.
6
Figure 1.3: Categorization of Face Recognition (FR) methods.
1.2.1. Face Recognition Processing
The face recognition processing flow, depicted in Figure 1.4, consists of four modules:
detection, alignment, feature extraction, and matching. Generally speaking, the whole
system modules or steps can be divided into two main stages: preprocessing and
recognition stages. The former stage includes face detection and alignment or localization
and normalization. Facial feature extraction and matching constitute the latter stage.
Face Recognition (FR)
Gray‐scale Image Color Image
Appearance based
Space Domain
PCA
Q ‐ PCA
ICA LDA
Feature based
Frequency Domain
Gabor Filters Wavelets
Preprocessing stage Recognition stage
Figure 1.4: Face recognition processing flow.
Face Detection ‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐
Tracking
Face Alignment
Feature Extraction
Feature Matching
Database of Enrolled
Users
Image
or Video
Face Location
Size & Pose
Aligned
Face
Feature
Vector
Face ID
7
Face detection segments the face areas from the background known as “nonface”
segments. Face alignment is used to achieve more accurate localization, and at
normalizing faces thereby it can treat the coarse estimates of the location and scale of
each detected face which is detected and provided by face detection module. Figure 1.5
illustrates face detection and alignment processes. The two processes yield the outlines of
located and normalized facial components outlines.
Feature extraction is performed to provide effective information that is useful for
distinguishing between faces of different persons and stable with respect to the
geometrical and photometrical variations. Face matching matches the extracted feature
vector of the input face against those of enrolled faces in the database. It is worth noting
that, the face recognition results depend highly on features that are extracted to represent
the face pattern and classification methods used to distinguish between faces. Also,
efficient distinct and effective features depend mainly on the face localization and
normalization processes.
Figure 1.5: Face detection and alignment processes.
8
1.3. Problem Statement
Human face trait has been proven to be a reliable candidate for person identification. As a
result, face recognition has become a very active research field. Consequently, several
research efforts have been employed for this active field. However, the great majority of
these efforts have used gray-scale representations of the face information. Although the
usage of gray-scale representation is considered as one of the more powerful methods for
face recognition, it has a major disadvantage of losing and discarding implicitly and
explicitly the valued color information. As a remedy to this drawback, recent approaches
have used the color information. However, each color component (i.e. red, green and
blue) is used individually. In this case, color independence is assumed which violates
basic principle of computer vision. Indeed, such approaches could perform better than
their gray-scale counterparts if they make use of the color information constructively.
1.4. Motivation
In this thesis, a solution based on hypercomplex (i.e., quaternionic) image representation
is proposed where the color information is used in a “holistic” manner. Hence, unlike
most of the existing approaches to handle the color information separately, the proposed
approach for face recognition implicitly accounts for the correlation that exists between
the color components. Figure 1.6, illustrates the proposed approach as an intersection of
four different areas.
9
1.5. Thesis Contributions
The major contributions of this research work are:
A novel scheme for efficient color face representation and processing for face-based
biometric recognition system.
The concept of principal component analysis (PCA) and singular value decomposition
(SVD) are extended to the quaternionic and carried over into the hypercomplex to
compactly represent color face images.
Detailed analysis of the incorporation of color information in face recognition is
provided.
Figure 1.6: Four areas of related work and the current research location.
10
Performance improvement by using color information and the Q-PCA method.
Extended performance evaluation is preformed and comparisons from different
perspectives are provided.
1.6. Thesis Outline
This thesis is organized as follows, Chapter 1 provides an introduction of biometric traits
and systems. Face recognition concepts and techniques are laid out. In Chapter 2, a set of
previous related works are presented and categorized into Grayscale-based and Color-
based face recognition approaches.
An overview of Quaternion (Hypercomplex) concepts is provided in Chapter 3. Basic
definitions, properties, and representations are discussed therein in details.. Chapter 4,
defines Eigenface method for face recognition. Besides, a brief background of principal
component analysis (PCA). Furthermore, a detailed description of the proposed
quaternionic principal component analysis (Q-PCA) method is presented along with
related mathematical expression of singular value decomposition (SVD) and it
quaternionic counterpart (QSVD).
In Chapter 5, the methodology experimental results are reported and discussed along with
a detailed performance analysis. A performance comparison with the current state-of-the-
art methods is carried out. Finally, Chapter 6 gives a summary of the thesis work along
with the contributions made in this thesis in light of the research and the testing results.
The chapter concludes with an outline of some proposed research directions where the
work described in this thesis can be further investigated.
11
CHAPTER 2
2. LITERATURE REVIEW
Biometrics is a very active area that has a lot of research efforts. Biometric systems
provide novel solutions to different access, control and security application. Face
recognition techniques evolved from being simple and limited to advanced, complicated,
and less limited.
2.1. Grayscale-based Face Recognition
Zhao et al. [24] provide a good comprehensive and critical face recognition literature
survey. Furthermore, relevant topics such as psychophysical studies, system evaluation,
and issues of illumination and pose variation are discussed.
2.1.1. Grayscale PCA-based Approach
Eigenfaces were first considered for face recognition by Turk and Pentland [22]. 2D gray-
face images are projected onto a feature space that captures the significant variation
among known face images. These significant features include the eigenfaces because they
are eigenvectors or principal components of the set of faces. The projection operation
characterizes an individual face by a weighted sum of the eigenface features. Hence in
order to recognize a particular face it is necessary only to compare these weights to those
of known individuals.
12
A method for gray-face recognition using principal component analysis (PCA) and radial
basis function (RBF) neural networks is presented by Thakur et al. [25]. PCA is employed
to reduce the dimensionality of the image and to keep some of the variations in the image
data. Adapted (RBF) neural networks are used under imposed conditions. Reported
experimental results show that the proposed method enhances the recognition
performance.
2.1.2. Grayscale Gabor Filters Approach
A novel algorithm for face recognition using Gabor features to train neural networks is
introduced by Rahman and Bhuiyan [26]. The algorithm is applied to gray morphed face
images within a constructed system using different scales and orientations. The Multi-
Layer Perceptron (MLP) neural network with back-propagation algorithm is employed for
face recognition and incorporates the convolution Gabor filter responses.
Wang et al. [27] propose an effective algorithm for face recognition using gray-face
Gabor image and Support Vector Machine (SVM) for face recognition. The proposed
approach derives the face Gabor image by down-sampling and concatenating the Gabor
wavelets representations. These Gabor wavelets are convolution of the face image with a
family of Gabor kernels, and then (2D-PCA) method is used to extract the feature space.
Then, the extracted features are fed to the SVM classifier.
13
2.1.3. Grayscale Wavelet-based Approach
In [28], Liu and Dai propose a facial representation based on the dual-tree complex
wavelet transform (DT-CWT). The proposed method effectively represents the structures
in gray facial image with low redundancy.
By exploiting the DT-CWT and independent component analysis (ICA), a novel face
recognition method is proposed by Chai et al. [29]. It provides a better representation for
feature extraction. Using PCA, the dimension of the DT-CWT feature vectors is further
reduced. Then, the ICA is exploited to reduce the feature redundancies and derive the
independent feature vectors.
2.2. Color-based Face Recognition
Several approaches make use of color information [6-8, 16, 20].
Choi et al. [6, 7] propose a metric called “variation ratio gain” (VRG) to theoretically
prove the significance of color effect on low-resolution faces. In addition, extensive
performance comparison studies are conducted.
Yang and Liu [8] introduced a General Discriminant Model (GDM) for color face
recognition. It is deals with two sets of variables: a set of color component combination
coefficients for color image representation and a set of projection basis vectors for image
discrimination. In order to find the optimal solution of the model an iterative whitening-
maximization (IWM) algorithm is designed and used.
14
An evolutionary two-fold framework is proposed by Shih and Liu [16]. First, two new
color spaces are defined as linear transformations of the input RGB (i.e., red, green and
blue) color space. The first color space is defined by one luminance (L) channel and two
chrominance channels (C1, C2). The second color space incorporates one luminance
channel (L) and three chrominance channels (C1, C2, C3). A genetic Algorithm (GA)
searches for the optimal transformations from the RGB color space to the (LC1C2) and
(LC1C2C3) color spaces, respectively.
Since color inputs have large dimensionality which increases the computational cost in FR
systems, it is important to determine the scenarios in which usage of color information
helps the FR system. Ganapathi [20] provides an empirical study for this purpose and
indicates the following observations: the inclusion of chromatic information in FR
systems is shown to be particularly advantageous in poor illumination conditions, a color
input of optimal dimensionality would improve the FR performance.
The diagonal nonnegative matrix factorization (BDNMF) is proposed by Wang et al. [3]
for color face representation and recognition. The approach employs block diagonal
matrix to encode color information of different channels. An adapted Nearest
Neighborhood classifier is used to identify color face samples. Whereas the obtained
experimental results on (CVL) and (CMU PIE) color face databases verify the
effectiveness of the proposed approach.
Thomas et al. [21] investigate the use of the characteristics of a 3D color to generate a
color Linear Discriminant Analysis (LDA) subspace, which could be used to recognize a
new testing image. In order to test the potential improvement in accuracy, a recognition
15
rate across two color face databases is computed. The observed results indicated that the
LDA color subspace significantly improves recognition accuracy over the standard
grayscale approach.
2.2.1. Color PCA-based Approach
The PCA technique is used repeatedly in various adaptations for color face recognition.
Actually, the algorithm based on PCA, form the basis of numerous studies in the
psychological and algorithmic face recognition literature as indicated in [12].
Wang et al. [2] propose a color face recognition approach based on 2D-PCA. The
proposed approach comprises a matrix-representation model that encodes the color
information directly to describe the color face image. In this way, color face images are
represented efficiently in the matrix format. Consequently, color-eigenfaces are computed
for feature extraction using 2D-PCA, and the Nearest Neighborhood classification is
adapted to identify the color face samples.
Moon and Phillips [12] introduce a generic modular PCA-algorithm in order to investigate
the influence of the computational and performance aspects of PCA-based for face
recognition algorithms.
2.2.2. Color Quaternion-based Approach
The first attempt of using the quaternion concept for color images is attributed to Pei and
Cheng [13]. They propose a quaternion model called the quaternion-moment block
16
truncation coding (QMBTC). It is proposed for compressing color-pixel blocks. It is
mainly based on color (BTC) algorithm and is derived by using the quaternion arithmetic
and moment preserving principle.
2.2.2.1. Color Quaternion PCA (Q-PCA) Approach
A technique for quaternion matrix algebra which can be used to process the eigen analysis
of a color image is proposed by Le Bihan and Sangwine [10]. This technique introduces
extensions of two classical techniques to their quaternionic case: Singular Value
Decomposition (SVD) and Karhunen-Loeve Transform (KLT). It also introduces the
problem of Eigne Value Decomposition (EVD) of quaternion matrix. The properties of
these quaternion tools are given and their behavior on natural color images is presented.
Furthermore, a method to compute the decomposition using complex matrix algebra is
provided. In addition, another consequent work by Sangwine and Le Bihan [31], the
Jacobi algorithm for computing the quaternion SVD (Q-SVD) is presented.
Based on Q-SVD, Shi [30] implementes the Q-PCA, and applies it to several applications
such as color image segmentation. Trilateral filtering is proposed by locally adapting
color and changing the shape of the filter to achieve the effect of smoothing colors that
preserve the edges.
Ding and Feng [11] propose another method of quaternion Karhunen-Loeve Transform
(Q-KLT) and biomimetic pattern recognition (BPR) for color face recognition. The model
of (Q-KLT) is used to extract the eigen-faces of training samples and algebraic features
for each by BPR method.
17
A method proposed by some researchers called Extended Two-Dimensional PCA (E2D-
PCA) which is an extension to the original 2D-PCA for gray face recognition. E2D-PCA
is promoted by a new covariance matrix preserving more local geometric structure
information than previous 2D-PCA methods. It avoids small sample size problem of the
PCA method. Although the new E2D-PCA improves the previous 2D-PCA, the former
one still has some limitation that, it is only considering the global information of face
images. Besides, it is only applied for gray face recognition. Where some local and color
information may be ignored. Thus, Chen et al. [15] provides a solution to such problems
by proposing SpE2D-PCA a hybrid approach based on sub-pattern technique and E2D-
PCA.
2.2.2.2. Color Quaternion Gabor Filters Approach
Numerous adapted biometric approaches by using quaternion concept with the well-
known Gabor-based methods are defined in [4] for color iris recognition and [1, 9, 23] for
color face recognition with distinct specifications of each.
A novel method for the automatic recognition and matching of color iris images is
proposed by Al-Qunaieer [4]. Particularly, the well-known IrisCode algorithm, developed
by Daugman for gray-scale iris images, is extended to the color domain by compact
quaternionic Gabor wavelets representation.
The Gabor filter, which is typically defined as a complex function, is extended to the
hypercomplex (quaternion) domain by Jones and Abbott [1, 9]. Several extensions to the
hypercomplex domain are discussed and a preferred formulation is selected.
18
Wei et al. [23] proposes a novel color face recognition method based on Local Binary
Patterns (LBP) of Quaternionic Gabor features (QGF). The approach is mainly based on
quaternion Gabor analysis within color image representation. This approach is
characterized by making full use of the interrelationship among different color channels to
enhance the performance of the face recognition system. (QGF) are used to encode the
positions and attributes of face elements. Then by using (LBP) which is a non-parametric
method was imposed on these (QGF) to obtain the robustness against variations of phase,
illumination, and facial expressions.
19
CHAPTER 3
3. QUATERNION (HYPERCOMPLEX) REPRESENTATION
3.1. Quaternion Numbers
Quaternions were first proposed in 1843 by W.R. Hamilton [36]. They are associative but
non-commutative and they belong to a specific class of hypercomplex numbers. Such
numbers are used in several applications such as mechanics in 3D space and 3D rotations
[4]. Also, quaternions are being applied and used in image processing to represent and
process color images in a holistic manner rather than each color separately. Complex
numbers represented as a combination of a real and an imaginary parts such as:
· (3.1)
where i is the imaginary unit, a and bi represent the real and imaginary parts, respectively.
Quaternions can be considered as a generalization of the complex numbers having one
real part and three imaginary parts [30]. A quaternion number can be written as a linear
combination defined as follows:
· · · (3.2)
where the quaternion q is made of one real part (a) and three imaginary parts (b, c, and
d),with i, j and k being the imaginary units. is the quaternionic set, is the complex set
and is the real set satisfying:
(3.3)
20
The imaginary parts i, j, and k satisfy the following properties:
ijk = i2 = j2 = k2 = −1 (3.4)
(ij = −ji = k), (jk = −kj = i), and (ki = −ik = j) (3.5)
3.2. Properties of Quaternion Numbers
Quaternions satisfy the multiplication rules which are sometimes known as Hamilton's
rules.
A quaternion with a zero real part is called a pure quaternion (i.e., a = 0):
0 (3.6)
The conjugate of a quaternion, q, is defined by negating its imaginary parts:
(3.7)
The modulus or magnitude of a quaternion, q, is defined by:
| | √ (3.8)
The addition or the sum of two quaternions, and is given by:
(3.9)
21
The multiplication or the product of two quaternions, and is given by:
(3.10)
The quaternion norm is defined by:
(3.11)
The norm is multiplicative such that:
· (3.12)
The division is uniquely defined except division by zero, therefore quaternions form a
division algebra. The inverse of a quaternion is defined by:
(3.13)
Given that the norm of q is non-zero.
To convert a quaternion to a regular complex number, where the quaternions can be
represented using its uniquely defined equivalent complex 2×2 matrices such that:
(3.14)
where z and w are complex numbers, a, b, c and d are real numbers, and and are
the complex conjugate of z and w, respectively.
22
3.3. Quaternion Representation of Color Face Images
Each pixel in a color image has the property of having the values of the three color
channels or components. These three colors simultaneously represent a pixel and are
defined as the three imaginary parts of the quaternion. Thus, any pixel or point, (x, y), of a
color face image can be represented by a pure quaternion as follows:
0 (3.15)
where x and y are the pixel coordinates, respectively. Figure 3.1 illustrates the approach
adopted to represent RGB color face images using a quaternionic form.
In this method, quaternions are used to represent the RGB color space. Therefore, the
three colors channels are processed equally in arithmetic and geometric operations such as
multiplication.
Figure 3.1: Quaternion representation of a color face image.
q = 0 + R.i + G.j + B.k
RGB image
Red channel Blue channel Green channel
23
3.4. Quaternion Fourier Transform (QFT)
The 2D Fourier Transform known as Fast Fourier Transform (FFT) is a useful image
processing method used to decompose a grayscale image into its sine and cosine
components. This method transforms a given image from spatial domain to Fourier or
frequency domain, where each pixel represents a specific frequency. It is employed in
several applications such as image analysis, filtering and compression. The 2D FFT can
given by [4]:
, , e e (3.16)
where I(u, v) is the frequency-domain representation of the image, and u and v are the
frequency coordinates, I(x, y) is the spatial-domain representation of the image, x and y
are the spatial coordinates.
While the FFT is applied on real and complex numbers, it can be extended to the QFT for
handling quaternionic numbers defined as [30]:
, e . , . e . (3.17)
Equation (3.17) can be generalized to obtain three different types of the QFT: the left-side
QFT, the right-side QFT and the two-side QFT. Moreover, The Inverse Quaternion
Fourier Transform (IQFT) can be performed for all types. The two-side QFT is defined as
[38]:
, e . , . e . (3.18)
24
where h(x, y) is the input Quaternion image, μ1 and μ2 are two pure quaternionic units,
orthogonal to each other, w and v are the spatial frequencies in x and y directions,
respectively.
The two-side IQFT is defined as:
, 14
e . , . e . (3.19)
The left-side QFT is defined as:
, e . , . (3.20)
The left-side IQFT is defined as:
,14
e . , . (3.21)
The right-side QFT is defined as:
, , . e . (3.22)
The right-side IQFT is defined as:
,14
, . e . (3.23)
25
3.5. Quaternion Gabor Filters
The Gabor filters are well known technique used for image characterization, texture
analysis and feature extraction [9]. The complex Gabor filters are widely used in the
literature, whereas very few research works make use of the extended Gabor filters in the
Quaternion domain. A 2D Gabor filter can be defined by [1]:
, 1
2exp
12
exp 2 (3.24)
where σx and σy are the space constants of the Gaussian envelope along the x and y axes.
They also characterize the bandwidth of the filter. Each different frequency is given by:
(3.25)
The Quaternion Gabor filters can be implemented by the QFT. Therefore, a 2D
Quaternion Gabor filter can be defined by an extension of Equation (3.25) as [1]:
, 1
2exp
1
2
2
2
2
2 exp 2 . exp 2 (3.26)
Another extension of Equation (3.25) can define a pure Quaternion Gabor filter is
achieved by multiplying the quaternion unit μ by the real part of the filter such as [1]:
, 1
2exp
12
cos 2 (3.27)
26
CHAPTER 4
4. PRINCIPAL COMPONENT ANALYSIS (PCA)
The PCA or KLT is a dimensionality reduction technique based on extracting the desired
number of principal components of the multidimensional data [14]. Another definition of
PCA is a general method of finding orthogonal axes that contain the most information
about a set of sample vectors in a multidimensional space [1].
PCA is the simplest of the true eigenvector-based multivariate analysis. It is considered as
a standard tool in modern data analysis in several fields including computer graphics. The
concept of PCA can be diagrammatically represented as shown in Figure 4.1. PCA is
usually related to the mathematical technique of SVD. A typical PCA mechanism can be
summarized in the following steps [17]:
1) Organize data as an m×n matrix, where m is the number of measurement types and
n is the number of samples.
2) Subtract off the mean form each row or column of the m×n matrix.
3) Apply SVD to calculate the eigenvectors of the covariance.
Figure 4.1: The PCA concept, (Li et al. [14]).
27
Suppose T= {t1,t2, … , tn}, is a training set of (facial) images consisting of n samples.
Each tk is an (i × j) image matrix representing the kth sample in the training set T. In PCA,
each tk is first reshaped into a high-dimensional vector, h (i ∗ j)×1 , by concatenating its
columns or its rows. Therefore, an (m × n) matrix, H, of all reshaped training samples is
obtained. m is the number of measurement types and n is the number of samples. Then,
the mean of whole training set, µ, is computed as follows [9]:
1
(4.1)
The covariance matrix, C, is computed as follows [17]:
1
(4.2)
4.1. Eigenfaces
Eigenfaces are a set of eigenvectors used in PCA-based human face recognition. The use
of eigenfaces for recognition was first proposed by Sirovich and Kirby in 1987 [22].
Informally, eigenfaces can be considered as a set of standardized face ingredients or
substances, derived from statistical analysis of many face images. The generated
eigenfaces can be represented visually as shown in Figure 4.2. In this case, a face image
can be considered as a combination of standard faces (eigenfaces).
28
4.2. Solving PCA Using SVD
The PCA calculation can be implemented via the SVD technique. The SVD of an (m × n)
matrix A (m ≥ n) is given by:
∑ (4.3)
where the m × n matrix U and the n × n matrix V have orthonormal columns. The n × n
matrix ∑ has the singular values of A with 0 along its main
diagonal and zero elsewhere. It should be noted that a singular value of a matrix, A, is the
square root of an eigenvalue of matrix AAT. Note that, the SVD allows for efficient and
robust computation of the PCA without the need to estimate the data covariance matrix.
4.3. Quaternion Principal Component Analysis (Q-PCA)
In the real domain, the eigenvalues and eigenvectors of real matrix can be easily
computed. However, computing those of a quaternion matrix is a difficult task [15]. In
order to calculate the PCA projection matrix based on a quaternion representation, it is
Figure 4.2: In the top gray-eigenfaces and in the bottom color-eigenfaces.
29
necessary to obtain the orthogonal eigenvector set of the covariance. The orthogonal
eigenvectors of the PCA covariance matrix can be obtained using three PCA-based
methods: Quaternion Singular Value Decomposition (Q-SVD), Quaternion Eigen-Value
Decomposition (Q-EVD), and Quaternion Karhunen-Loeve Transform (Q-KLT).
The Q-SVD can be considered as a generalization of the real or complex SVD. It inherits
similar properties [30]. The decomposition of a quaternion matrix by means of classical
complex algorithms is based on the SVD computation of the complex adjoint matrix
(CAM). Therefore, in the Q-SVD solution, a quaternion matrix can be decomposed into
its singular value form by converting it into complex matrix representation. Each
quaternion matrix of size M×N has an equivalent complex matrix of size 2M×2N
[30, 31]. Hence, the existing complex SVD algorithm can be applied to this equivalent
complex matrix to obtain the eigenvectors and eigenvalues (i.e. singular values) of the
corresponding quaternion matrix .
4.3.1. Quaternion Singular Value Decomposition (Q-SVD)
The Q-PCA calculation can be implemented via the Q-SVD technique. Based on a
theorem given in [34] to show the existence of the SVD of a quaternion matrix. Let
be a n×n quaternion valued matrix with rank k, and there exist two unitary
quaternion matrices and such that:
· · ∑ 00 0
(4.4)
30
where ( ) represents the Hermitian transpose (or conjugate-transposition) operator. Note
that, ( q , c and r) denote data types (quaternion, complex and real), respectively. The
unitary quaternion matrices and have the property such that:
· · (4.5)
Hence, the multiplication of quaternion matrices actually yields the real identity matrix
. The notation ∑ is a real diagonal matrix ( ∑ ) (i.e., the number of non-null
singular values), k is the rank of . The values on the diagonal of ∑ ,… ,
with (1 ), are the real positive singular values of the quaternion matrix ,
arranged in decreasing magnitude order along the diagonal. Thus, Equation (4.5) can be
rewritten as:
· ∑ 00 0
· (4.6)
According to a definition of the equivalent complex matrix of a quaternion matrix [34].
Each quaternion defined by:
· · · (4.7)
This quaternion q can be decomposed using:
· · · · (4.8)
where a and b are two complex numbers.
For instance, an n×n quaternion valued matrix defined as:
· (4.9)
31
where A and B are n×n complex matrices, then from equation (3.14) in Chapter 3 and
equation (4.9) the equivalent 2nx2n complex matrix of is
(4.10)
Thus, the classical complex SVD algorithm can be applied to to generate the eigen-
basis (i.e., eigenvectors) and the corresponding singular values ranked in descending
order.
Let the SVD of a quaternion matrix be:
· ∑ · (4.11)
Let the equivalent complex matrix for be:
· ∑ · (4.12)
then,
∑ ∑ (4.13)
If 2
2
(4.14)
Then, ·
·(4.15)
32
where, means the odd rows and means the odd columns of matrix
M. In addition, some of the most significant properties of the Q-SVD, when applied to
color images, are listed below by Le Bihan et al. [10]:
1) Invariance to spatial rotation (also true in the case of greyscale images with SVD).
2) Invariance to spatial shift (vectors in U and V are shifted by the same amount).
3) Invariance to color space rotation.
4.3.2. QSVD-Based Color Image Compression
Several useful methods applied to the SVD can be extended to color images using Q-
SVD. Such extensions are achieved without separating the color image into three channels
and processing each color channel independently. One very useful SVD-based method is
image compression which definitely can be extended to color images. The robustness of
this compression method is in storing large images as smaller more manageable ones.
QSVD-based compression accomplished by reproducing the original image with each
succeeding non-zero singular value ( ∑). Furthermore, by using fewer singular values
( ′ ) it can achieve more compression to produce approximate image [30]. Using equation
(4.12), the image can be decomposed as:
· ∑ · (4.16)
Then, to construct the eigen-image of original one, the SVD of a color image, Q, can be
decomposed into:
33
· ∑ · (4.17)
where ui and vi are the column vectors of matrix U and V, respectively. The λi’s are the
diagonal elements of the real matrix ∑ and R is the rank of Q. Here, each i product
generates an eigen-image. Therefore, the constructed color image, x, can be
considered as the linear combination of R color eigen-images. Note that, in quaternion and
complex matrices, the preceding eigen-images represent the low-frequency components of
the original image, and the later ones represent the high-frequency components [30].
Therefore, an approximate of a color image can be obtained by summing the first k
eigen-images such that:
(4.18)
where the real part of is small, and decreases to zero when k increases to R.
However, a good approximation of the original color image can be provided by a small k.
Due to this compression, the storage requirements for the color image is reduced from
(3×N×N) to k(2*4N+1).
Figure 4.3 illustrates four estimated images based on Q-PCA and its approximation
reconstruction applied to a facial image of size 213×144. Note that when the image has
more high-frequency components, a larger k is necessary to have a good approximation.
Otherwise, for images containing mostly low-frequency signals, small k is good for the
reconstruction. As shown in the two approximations in (c) and (d), it is clear that a
satisfactory approximate representation can be found with far fewer singular values. This
34
simple example illustrates the power and efficiency of the Q-SVD method. The
reconstructed image is produced by the first 16 singular values requiring only
16×213×4=13,632 entries. Another major advantage is performance optimization of a
recognition system.
(a) (b) (c) (d)
Figure 4.3: Q-SVD-based image compression. (a),(b),(c) and (d) are the reconstructed images with k=8,16,25,144 respectively. Where (d) is the perfect reconstruction of the original image
4.3.3. Proposed Q-PCA Algorithm
The most significant concepts (i.e., Quaternion, PCA, SVD) associated with the novel
proposed scheme have been defined previously. The proposed QPCA algorithm, called
Quaternion Principle Component Analysis, can be described and summarized under a list
of steps needed to carry out Q-PCA procedure as follows:
1) For each facial color image in the training data, convert into pure
quaternionic representation.
0 · · · (4.19)
35
2) Organize the data as an (m×n) design matrix, , where m is the number of
measurement types (also rows) and n is the number of samples (also columns).
Each matrix cell mxny is a quaternion number. Hence, each column ny represents a
single vector or orthogonal representation of quaternion image pixels.
3) Subtract off the mean, , and divide by the variance, , each measurement type,
mx, to obtain n orthonormal vectors of the design matrix .
4) Calculate the Q-SVD or the quaternion eigenvectors, , of the covariance matrix,
defined as:
· ∑ · (4.20)
5) To recognize or match a query color face image, fq, the quaternion form of fq is
normalized and projected onto the eigen-face space by:
· (4.21)
where is the projection of the image , and ( ) represents the Hermitian
transpose (or conjugate-transposition) of the eigenvectors, . and are the
mean and variance vectors of the design matrix, , respectively.
36
CHAPTER 5
5. METHODOLOGY AND EXPERIMENT
5.1. Research Methodology
This research proposes a novel scheme for color face recognition. Particularly, the
proposed scheme is based on the well-known PCA method. In this research, the PCA is
naturally extended to the hypercomplex (quaternionic) domain to define the Q-PCA.
Besides the benefit of combining the existing color information, Q-PCA supports and
emphasizes the usage of other beneficial information (such as color correlation and
interaction) that would not be obtained using a mere parallel (and independent) processing
of the color components. The added-value brought in by the proposed scheme is made
possible by the holistic representation of the color information in the hypercomplex
domain. Moreover, a comparative analysis is carried out to highlight the improvement
attributed to the holistic processing of the color components. The experimental work is
conducted using standard color face databases.
5.2. Color Face Image Database
Within the experimental work of this thesis, two standard databases are used. The first one
is used as a minor database and the other one is used as major. The minor database is the
“Georgia Tech Face Database (GTDB)” which is used at earlier stages of the development
of the PCA technique for gray-level and color face images. The major database is “The
Color FERET Database” which is used in all experiments reported therein.
37
5.2.1. Georgia Tech Face Database (GTDB)
The GTDB was collected and prepared at the Center for Signal and Image Processing at
Georgia Institute of Technology. It contains images of 50 people taken in two or three
sessions between 06/01/1999 and 11/15/1999 to take into account the variations in
illumination conditions, facial expression and appearance. Furthermore, the faces were
captured at different scales and orientations. Each subject (a person) in the database is
represented by 15 color JPEG images with cluttered background taken at a resolution of
640×480 pixels. Hence, the database contains 750 face images. The average image is
150×150 pixels. The GTDB also provides a preprocessed set (GTDB_crop) of cropped
and relabeled images where each in the database was cropped and background sremoved.
An addition at preprocessing step is performed in this thesis. Due to the various sizes of
the cropped images, each image is resized to 213×144 pixels.
5.2.2. Color FERET Database
The FERET database is a well known standard facial database in both color and grayscale
representations provided by the US National Institute of Standards and Technology
(NIST). It has been designed to advance the state of the art in face recognition. The
database collection was a collaborative effort between Dr. Wechsler and Dr. Phillips. The
database was collected at various angles in 15 sessions between August 1993 and July
1996. It contains 1564 sets of images for a total of 14,126 images that includes 1199
individuals and 365 duplicate sets of images. A duplicate set is a second set of images of a
person already in the database and was usually taken on a different day. For some
38
individuals, two years have elapsed between their first and last sittings, with some
subjects being photographed multiple times. This period of time was important to enable
researchers to study the effects of changes in a subject's appearance that occur over years.
Thus, the FERET database introduces variability by the inclusion of images taken at
different dates and locations. This results in changes in lighting, scale and background.
The whole database is divided into a development set, provided to researchers, and a set
of sequestered images for testing. The images in the development set are used as a
representative of the sequestered images. However, the available database for researchers
is a subset of the database consisting of 11,338 images from the total of 14,126. These
images represent 994 individuals from the total of 1199 individuals. The remaining
individuals and images were not provided for the researchers. The database, provided in
color PPM-format, represents images in three different resolutions:
1) Full size images 768×512 pixels.
2) Half size images 384×256 pixels.
3) Quarter size images 192×128 pixels.
There are 13 different poses in the collected database representing the person’s face at
various angles. Table 5.1 gives a summary of the database content and the imagery pose
types.
39
Two letter code
Pose Angle (degrees) Description
Number of images in Database
Number of Subjects (persons)
fa 0 = Regular facial expression 1364 994 fb 0 Alternative facial expression 1358 993 ql -22.5 quarter left - head turned about 22.5 degrees left 761 501 qr +22.5 quarter right - head turned about 22.5 degrees 761 501 hl -67.5 half left - head turned about 67.5 degrees left 1267 917 hr +67.5 half right - head turned about 67.5 degrees right 1320 953 pl -90 Profile left - head turned about 90 degrees left 1312 960 pr +90 Profile right - head turned about 90 degrees 1363 994 ra -45 random image - head turned about 45 degree 321 261 rb -15 random image - head turned about 15 degree 321 261 rc +15 random image - head turned about 15 degree 610 423 rd +45 random image - head turned about 45 degree 290 236 re +75 random image - head turned about 75 degree 290 236
Table 5.1: Summary of FERET color face database.
5.3. Performance Evaluation
A biometric evaluation protocol determines how to test a system, select the data, and
measure the performance. In order to apply validation and verification to the proposed
approach, the (FERET) evaluation protocol is used, which is a standard method for
evaluating face-recognition algorithms [32]. This biometric evaluation protocol
investigates the performance of Q-PCA-based versus the typical PCA-based algorithms.
The PCA-based algorithms considered in this thesis are: gray-PCA, Red-component-PCA,
Green-component-PCA, Blue-component-PCA, Avg-RGB-PCA of the three components,
and SCFT-CPCA (complex PCA of Spatio-chromatic Fourier transform) [37].
A number of stress test experiments are performed for these different implementations to
evaluate and compare their performance. All these algorithms are examined by applying a
number of prober performance evaluation methods and using the same data sets. Table 5.2
shows the PCA-based algorithms used in this work.
40
Algorithm Description PCA Calculates PCA for grayscales face images QPCA Calculates Quaternion PCA for color face images Red PCA Calculates PCA for the Red component of color face images Green PCA Calculates PCA for the Green component of color face images Blue PCA Calculates PCA for the Blue component of color face images
Avg RGB PCA Calculates PCA for the Average of the three color channels (i.e. red, green and blue) of color face images
SCFT CPCA Calculates Complex PCA for the spatio-chromatic Fourier transform of color face images
Table 5.2: Summary of PCA-based algorithms used in this thesis.
5.3.1. Standard Testing Subsets
In this experimental work, FERET standard testing subsets are used to carry out the
performance and statistical evaluation. Particularly, the gallery and probe sets used in
FERET tests (September 1996) are selected for this purpose. Therefore, the identification
scores are carried out for four categories of probes and all the tests for these probes are
using a single gallery containing 993 images as indicated in Table 5.3.
Evaluation Task Recognized Names Gallery (993) Probe Set Facial Expression FB gallery.names probe_fafb_*.names (993) Low aging of subjects duplicate I gallery.names probe_dup_1_*.names Illumination fc gallery.names probe_fafc_*.names (98) High aging of subjects duplicate II gallery.names probe_dup_2_*.names
Table 5.3: Gallery and four probe sets.
Table 5.4 gives the identification scores derived two times for the same gallery and probes
in two different versions (or resolutions). This redundancy aims at assessing the effect of
image resolution on the system performance.
Dataset No. of images / Resolution version Gallery (993) / 192×128 pix (993) / 384×256 pix FB (probe) (993) / 192×128 pix (993) / 384×256 pix duplicate I (probe) (736) / 192×128 pix (736) / 384×256 pix fc (probe) (98) / 192×128 pix (98) / 384×256 pix duplicate II (probe) (228) / 192×128 pix (228) / 384×256 pix
Table 5.4: Resolution of gallery and probe sets.
41
The first probe category is the FB probes. In each set of images, there are two frontal
images denoted by fa and fb pose types. One of the images is randomly placed in the
gallery, and the other image is placed in the FB probe set. This category is aimed to
evaluate the algorithm performance against the variation in facial expressions. The second
probe category is the duplicate I probe, which contains all duplicate frontal images in the
FERET database for the gallery images. This category is aimed to evaluate the algorithm
performance against the variation in low degree of subject aging. The third probe category
is the fc probe, where images are taken the same day with a different camera and lighting.
This category is aimed to evaluate the algorithm performance against the variation in
illumination. The fourth probe category is the duplicate II probe, which consistes of
duplicates where there is at least one year between the acquisition of the probe image and
the corresponding image in the gallery. This category is aimed to evaluate the algorithm
performance against the variation in high degree of subject aging. Figure 5.1 gives the
face images of one subject samples from the gallery and all the four probes.
Another subpart of color FERET database consists of five subsets. Each subset has its
own gallery and probes (one gallery and two probes). The gallery of each subset contains
approximately 200 individuals. For each subset, one of the two probes represents the FB
fa fb duplicate I fc duplicate II
Figure 5.1: One subject face images (gallery and probes).
42
category and the other represents the duplicate I category. Table 5.5 lists the five subsets
with their details.
Probe Name Gallery size FB Probe set size duplicate I Probe set size Set 1 200 200 145 Set 2 200 200 70 Set 3 200 200 204 Set 4 200 200 283 Set 5 193 193 34
Table 5.5: Subsets used to investigate the variation in performance.
5.3.2. Test Design
The adopted biometric evaluation protocol investigates the performance aspects to several
FR algorithms. In this testing protocol, each evaluated algorithm can use a different
similarity measure and it does not compare similarity measures from different algorithms.
However in the current experimental work the similarity measure is the same for all
compared algorithms, since they are all PCA-based algorithms. A significant advantage of
this protocol is that for any two images, qi and tk, the similarity measure, si(k), is known
which allows for a greater flexibility and more comprehensive evaluation methodology.
This flexibility is achieved and satisfied by computing scores for virtual galleries and
probe sets. The algorithm performance is characterized by different categories of images:
1) Rotated images.
2) Duplicates taken within a week of the gallery image.
3) Duplicates where the time between the images is at least one year.
4) Galleries containing one image per person.
5) Galleries containing duplicate images of the same person.
43
A gallery G is a virtual gallery if G is a subset of the target set T . Similarly, P is a virtual
probe set if P is a subset of query set Q. For a given gallery G and probe set P, the
performance scores are computed by examination of similarity measures si(k) such that qi
P and tk G.
In all the conducted experiments, for all Algorithms the number of eigenvectors is
adopted to be 200 eigenvectors. This decision is obtained after performing a number of
trials using 25, 50 and 100 eigenvectors. Selecting this certain number of eigenvectors to
be used for projecting a tested facial image onto the eigenface space improves the
accuracy in general about 3-9%.
5.3.3. Standard Performance Measures
For the sake of comparison with existing algorithms, a set of standard biometric
evaluation methods are provided and described in this section. Such evaluation methods
or measures are commonly used to allow a fair comparison and a precise statistical
evaluation.
Cumulative Match Characteristic (CMC): It is usually computed and visually
represented as a curve of ranked cumulative match scores form rank of 1 to rank of k ≤
the number of enrolled persons.
In order to define the decision theory of CMC measure for evaluating the performance of
an algorithm, there exist two basic models are: the closed and open universes. In the
closed universe, every probe is in the gallery whereas in an open universe, some probes
44
are not in the gallery. Since the concern of this research is concerned with identification,
the appropriate model is the closed-universe model which reports different performance
statistics and reflects important aspects of FR algorithms. Such aspects allow evaluators to
ask how good an algorithm is at identifying a probe image. Note that, the question is not
always “is the top match correct?” but it could be “is the correct answer in the top n
matches?”. The answer of such question allows evaluators to know how many images
(what are the k-images) have to be examined to get a desired level of performance. The
reported performance statistics are based on the CMC curve wherein the rank is plotted
along the horizontal axis and the percentage of correct matches is plotted in the vertical
axis. The cumulative match score can be calculated for any algorithm against any subset
of the probe set. The computation of an identification score as follows:
Let | P | be the size of probe P. The probe set P is scored against gallery G, where G =
{g1, …, gm} and P = {p1, …, pn} by comparing the similarity scores si(*) such that pi in P
and gk in G, where * indicates the similarities of all gk images in a gallery set, G, to be
computed against each pi in P. For each probe image, pi in P, si(*) is scored for all
gallery images gk in G. A smaller similarity score implies a closer match. If gk and pi are
the same image, then si(k) = 0. The function id(i) gives the index of the gallery image of
the person in probe pi, i.e., pi is an image of the person in gid(i). A probe pi is correctly
identified if si(id(i)) is the smallest scores for gk G. A probe pi is in the top k if si(id(i))
is one of the k-th smallest score si(*) for gallery G. Let Rk denotes the number of probes
in the top k. Thus, the fraction of probes in the top k is reported, such that Rk / | P |. For
example, let k = 5, R5 = 80 and | P | = 100. Based on this formula, the performance score
45
for R5 is 80/100 = 0.8. The design scheme of the testing procedure, used in this research is
illustrated in Figure 5.2.
In all reported performance results, the size of the gallery is the number of different faces
(people) contained in the images that are in the gallery. Hence, there is one image per
person in the gallery. Therefore, the size of the gallery is also the number of images in this
gallery. The number of probes scored is also the size of the probe set (i.e., | P | ).
However, the probe set may contain more than one image of a person and the probe set
may not contain an image of everyone in the gallery. Every image in the probe set has a
corresponding image in the gallery (closed-universe model).
Genuine distribution (GD): A distribution which is estimated by comparing different
face image samples that belong to the same person against each other.
It should be noted that there is no need to inverse the comparisons between any two
samples i and j since the comparison is symmetric.
Figure 5.2: Design scheme of the testing procedure.
PCA‐based Algorithm for Face Recognition
Training code
load
gi
Gallery images (G)
(One image/Person)
0001
0002 . . .
NNNN
List of probes (P)
0001
0002 . . .
NNNN
Test code
load
pi
match
gi and pi
Scoring code
Compute
performance score
Rk / |P |
finding Similarity
of each si(k) then si(id(i)) for a defined rank size
score
Rk
Results
46
Imposter distribution (ID): A distribution which is estimated by comparing different
face images that belong to different persons.
Such that, the ith sample of each person is compared to the ith sample of all the remaining
persons in the dataset. In this case, the comparison is also symmetric.
For illustration, a sample of GD and ID distributions is given in Figure 5.3. The more
separated are the GD and ID curves, the better is the matching process. Furthermore, the
intersection areas represent errors in the matching process.
Within this error area, two subareas exist:
False Accept Rate (FAR): A subarea which is the region of ID that is considered to
belong to the GD.
False Reject Rate (FRR): A subarea which is the portion of GD that is considered as part
of the ID.
Both error areas are illustrated in Figure 5.3. Consequently, the selection of a specific
threshold represents a trade-off between the FAR and FRR errors. The selection of a
proper threshold depends significantly on the application of the biometric system at hand.
For instance, the FAR error is selected to be large in forensic applications for criminal
identification to reduce the possibility of missing a wanted criminal. On other hand, the
FRR error is selected to be large in highly secure access control systems to guarantee that
authorized genuine users can gain access.
47
Generally speaking, better results are achieved if the two distributions are well separated.
Decidability index (d): The separation or the distance between the two means the GD and
ID considered as a hint about the system performance.
The decidability index, d, can be calculated as follows [4]:
(5.1)
where µ and µ are the means of the GD and ID, respectively. Large values for the
decidability index do not necessarily mean that the performance is better. In some cases,
although the decidability index is large, the errors can be large and should be taken into
account with the decidability index.
The FAR and FRR error curves can be calculated from the GD and ID by using a
threshold t that ranges from 0 to 1. Thus, the FAR(t) and FRR(t) are functions of t, where
FAR (t) is the percentage of imposters greater than or equal to the threshold, and FRR(t)
is the percentage of genuine less than the threshold [35]. Besides, the following
information can be estimated from the FAR and FRR curves as illustrated in Figure 5.4:
Figure 5.3: GD versus ID and FAR versus FRR errors.
Threshold(t)
FAR
1
FRR
0
p
Matching score
Imposter distribution
Genuine distribution
48
Equal-Error Rate (EER): The error rate where accept and reject errors are equal
(FAR(t) = FRR(t)), so that lower EER values indicate less errors.
ZeroFRR: The lowest FAR where no FRR occurs.
ZeroFAR: The lowest FRR where no FAR occurs.
Receiver Operating Characteristic (ROC): A curve is used to measure and summarize
the performance of a biometric identification or verification systems.
An ROC curve is based on the representation of the decision thresholds, therefore it
shows the system performance at all thresholds. Note that it is threshold independent,
allowing performance comparison of different systems under similar conditions. It plots
the percentage of impostor attempts accepted which is the FAR on the x-axis, against the
percentage of genuine attempts accepted which known as correct acceptance rate (CAR=1
- FRR) on the y-axis [35]. In ROC curves the relationship between FAR and FRR can be
observed. Such that, as the FAR error is increased, the FRR is decreased and accordingly
FRR(t) FAR(t)
ZeroFAR
EER
t
erro
r
ZeroFRR
Figure 5.4: FAR versus FRR curves.
49
the CAR or (1- FRR) is increased, and vice versa. The best curve is a straight horizontal
line with zero FRR that means accordingly the genuine attempts acceptance is (1-0=100),
which means that EER is equal to zero. Consequently, the lower curve means more errors.
5.3.4. Identification Performance against Probe Categories
The identification performance is investigated and reported for each of the compared
algorithms against four different probe categories from the first standard dataset. This
identification performance evaluation is based on a fraction of CMC scores ranked from 1
to 50 for the sake of comparison. Moreover, the identification performance for an
algorithm is investigated and reported for the same dataset in different resolutions. Table
5.6 gives a summary of the properties of the test gallery and probes.
Figure No. Probe Category Gallery size Probe set size Resolution
5.5 and 5.6 FB 993 993 19
2×12
8 pi
x
384×
256
pix
5.7 and 5.8 duplicate I 993 736
5.9 and 5.10 fc 993 98
5.11 and 5.12 duplicate II 993 228
Table 5.6: Summary of used gallery and probe categories.
Table 5.7 gives the fraction of probes whose gallery match is top ranked (i.e., the rank
equals to 1). Hence, each number in this table represents the correct total match score
considering the match with only one nearest-neighbor to each tested sample. This is
effectively the performance on forced identification (i.e., best guess). As shown in the
table, each probe category has two score numbers representing scores of two different
resolutions.
50
Gallery Size / Scored Probes 993 /993 993 /736 993 /98 993 /228
FB duplicate I fc duplicate II Algorithm 192×128 384×256 192×128 384×256 192×128 384×256 192×128 384×256
PCA 0.624 0.626 0.135 0.132 0.010 0.010 0.004 0.004 QPCA 0.646 0.648 0.145 0.147 0.010 0.010 0.004 0.004 Red PCA 0.635 0.638 0.141 0.143 0.010 0.010 0.004 0.004 Green PCA 0.621 0.622 0.130 0.132 0.010 0.010 0.004 0.004 Blue PCA 0.599 0.600 0.128 0.129 0.010 0.010 0.004 0.004 Avg RGB PCA 0.619 0.620 0.133 0.135 0.010 0.010 0.004 0.004 SCFT CPCA 0.004 0.002 0.001 0.004 0.000 0.000 0.000 0.000
Table 5.7: Match scores of the top-rank for each PCA-based algorithm.
Table 5.8 gives the estimated average score of the total match score against each probes
(i.e., the average of scores located along a rank from 1 to 50) which consider the average
of the first 50 identification estimations. Hence, each number in this table represents the
average of correct total match scores of the ranks from 1 to 50 (i.e., from 1 to 50 nearest-
neighbors) for each test sample. This is effectively the total performance on forced
identification (i.e., best 1,2, …, 50 guesses). As shown in the table, each probe category
has two score numbers representing scores of two different resolutions.
Gallery Size / Scored Probes 993 /993 993 /736 993 /98 993 /228
FB duplicate I fc duplicate II Algorithm 192×128 384×256 192×128 384×256 192×128 384×256 192×128 384×256
PCA 0.867 0.868 0.277 0.277 0.037 0.037 0.047 0.047 QPCA 0.880 0.881 0.284 0.285 0.042 0.041 0.052 0.052 Red PCA 0.887 0.887 0.290 0.290 0.030 0.029 0.035 0.035 Green PCA 0.862 0.862 0.273 0.274 0.045 0.045 0.054 0.054 Blue PCA 0.846 0.846 0.269 0.269 0.056 0.057 0.060 0.061 Avg RGB PCA 0.865 0.865 0.277 0.278 0.044 0.044 0.050 0.050 SCFT CPCA 0.059 0.047 0.025 0.035 0.017 0.035 0.020 0.030
Table 5.8: Average score of the total match scores of rank 50 for each PCA-based algorithm.
In order to evaluate and compare the performance of different FR algorithms, two aspects
are taken in account. The first is the evaluation of an algorithm based on the achieved top-
rank match score as reported in Table 5.7. The second is the total-average of total match
scores that depends on averaging the total of scores achieved as listed in Table 5.8. So,
51
according to the scoring results summarized in Tables Table 5.7-Table 5.8, to a number of
observations can be estimated.
By testing the PCA-based algorithms against FB probe data, as illustrated in Figures
Figure 5.5-Figure 5.6, the Q-PCA algorithm yielded the highest top-rank match score, and
achieved the best performance in the identification task. While the Red PCA achieved the
second best top-rank score, it achieved a total-average score better than the Q-PCA
algorithm. It should be noted that, the total-average score of each algorithm was not
affected by the probe resolutions except in the case of the 384×256 pixels resolution,
where the top-rank score of the SCFT CPCA algorithm is slightly decreased.
The performance of all algorithms against duplicate I probe data with 192×128 pixels
resolution is illustrated in Figure 5.7. The best algorithm has the highest top-rank score
and the worst algorithm has the lowest top-rank score. It can be seen that the Q-PCA
algorithm achieved the best top-rank and the SCFT CPCA algorithm achieved the lowest
top-rank. On the other hand, the performance was quite differed with respect to the total-
average. The best total-average was achieved by the Red PCA algorithm. The Q-PCA
algorithm was the second. However, against the duplicate I probe data using 384×256
pixels resolution, as illustrated in Figure 5.8 the performance trend has slightly changed.
According to the reported top-rank scores the Q-PCA algorithm has still the best
performance. From the total-average perspective, the performance has improved for the
Q-PCA, Green PCA and Avg RGB PCA algorithms. The best total-average was achieved
by the Red PCA algorithm followed by the Q-PCA algorithm.
52
Figures Figure 5.9 -Figure 5.10 illustrate the performance of all algorithms against the fc
probe data at resolutions of 192×128 pixels and 384×256 pixels, respectively. The top-
rank score value was the same for all algorithms in both resolutions as listed in Table 5.7
except for the SCFT CPCA algorithm. However, the total-average varies and can be used
for comparing and ranking the competitor algorithms. Unlike in the FB and duplicate I
datasets the best total-average is achieved by the Blue PCA algorithm. Then, the Green
PCA and Avg RGB PCA algorithms achieved the next highest performance followed by
the Q-PCA algorithm.
As illustrated in Figures Figure 5.11 -Figure 5.12, the performance of all algorithms against
the duplicate II probe data at two resolutions of 192×128 pixels and 384×256 pixels,
respectively. The top-rank score was the same for all algorithms in both resolutions as
listed in Table 5.7 except for SCFT CPCA. However, the total-average varies for each
algorithm. With the two resolutions, the best total-average was achieved by the Blue PCA
algorithm followed by the Green PCA algorithm.
53
Figure 5.5: Identification performance using the FB with 192×128 pixels resolution.
Figure 5.6: Identification performance using the FB with 384×256 pixels resolution.
54
Figure 5.7: Identification performance using the duplicate I with 192×128 pixels resolution.
Figure 5.8: Identification performance using the duplicate I with 384×256 pixels resolution.
55
Figure 5.9: Identification performance using the fc with 192×128 pixels resolution.
Figure 5.10: Identification performance using the fc with 384×256 pixels resolution.
56
Figure 5.11: Identification performance using the duplicate II with 192×128 pixels resolution.
Figure 5.12: Identification performance using the duplicate II with 384×256 pixels resolution.
57
5.3.5. Performance Comparison Using CMC in Different Resolutions
The effect of image resolution on performance is investigated and reported in this
experiment by comparing the CMC curves of each algorithm using the same probe
categories in different resolution. Table 5.9 shows, for each PCA-based algorithm, the
average score of the total match scores of the full rank applied to the four probe categories
using two resolutions. It should be noted that the bolded numbers in the table represent the
affected performance, either positively or negatively, by increasing the resolution.
Gallery Size / Scored Probes 993 /993 993 /736 993 /98 993 /228
FB duplicate I fc duplicate II Algorithm 192×128 384×256 192×128 384×256 192×128 384×256 192×128 384×256
PCA 0.987 0.987 0.612 0.613 0.086 0.085 0.193 0.194 QPCA 0.988 0.988 0.609 0.610 0.089 0.089 0.191 0.191 Red PCA 0.989 0.989 0.624 0.624 0.074 0.074 0.187 0.187 Green PCA 0.986 0.986 0.606 0.606 0.090 0.090 0.189 0.189 Blue PCA 0.981 0.981 0.588 0.589 0.112 0.112 0.186 0.187 Avg RGB PCA 0.985 0.985 0.606 0.606 0.092 0.092 0.187 0.188 SCFT CPCA 0.550 0.522 0.350 0.377 0.034 0.052 0.093 0.121
Table 5.9: Average score of the full rank for each PCA-based algorithm in two resolutions.
The comparison graphs for each algorithm are illustrated to represent the identification
performance curves using each probe category in two resolutions. The illustrated graphs
represent the full CMC with rank from 1 to the full rank size which is equals to full size of
the probe set.
58
5.3.5.1. CMC for the Gray-PCA algorithm
Figure 5.13: The CMC curves for PCA using the FB probes.
Figure 5.14: The CMC curve for PCA using the duplicate I probes.
59
Figure 5.15: The CMC curves for PCA using the fc probes.
Figure 5.16: The CMC curves for PCA using the duplicate II probes.
60
5.3.5.2. CMC for the QPCA algorithm
Figure 5.17: The CMC curves for QPCA using the FB probes.
Figure 5.18: The CMC curves for QPCA using the duplicate I probes.
61
Figure 5.19: The CMC curves for QPCA using the fc probes.
Figure 5.20: The CMC curves for QPCA using the duplicate II probes.
62
5.3.5.3. CMC for the Red-PCA algorithm
Figure 5.21: The CMC curves for Red-PCA using the FB probes.
Figure 5.22: The CMC curves for Red-PCA using the duplicate I probes.
63
Figure 5.23: The CMC curves for Red-PCA using the fc probes.
Figure 5.24: The CMC curves for Red-PCA using the duplicate II probes.
64
5.3.5.4. CMC for the Green-PCA algorithm
Figure 5.25: The CMC curves for Green-PCA using the FB probes.
Figure 5.26: The CMC curves for Green-PCA using the duplicate I probes.
65
Figure 5.27: The CMC curves for Green-PCA using the fc probes.
Figure 5.28: The CMC curves for Green-PCA using the duplicate II probes.
66
5.3.5.5. CMC for the Blue-PCA algorithm
Figure 5.29: The CMC curves for Blue-PCA using the FB probes.
Figure 5.30: The CMC curves for Blue-PCA using the duplicate I probes.
67
Figure 5.31: The CMC curves for Blue-PCA using the fc probes.
Figure 5.32: The CMC curves for Blue-PCA using the duplicate II probes.
68
5.3.5.6. CMC for the Avg RGB-PCA algorithm
Figure 5.33: The CMC curves for Avg RGB-PCA using the FB probes.
Figure 5.34: The CMC curves for Avg RGB-PCA using the duplicate I probes.
69
Figure 5.35: The CMC curves for Avg RGB-PCA using the fc probes.
Figure 5.36: The CMC curves for Avg RGB-PCA using the duplicate II probes.
70
5.3.5.7. CMC for the SCFT-CPCA algorithm
Figure 5.37: The CMC curves for SCFT-CPCA using the FB probes.
Figure 5.38: The CMC curves for SCFT-CPCA using the duplicate I probes.
71
Figure 5.39: The CMC curves for SCFT-CPCA using the fc probes.
Figure 5.40: The CMC curves for SCFT-CPCA using the duplicate II probes.
72
5.3.6. Variation in Identification Performance
Another performance evaluation approach aims to investigate the identification
performance using different galleries that contain different face images from the other
galleries. While a face-recognition algorithm estimates the identity of a face, a possible
question “how does the algorithm performance change using a different gallery and probe
set?” Tables Table 5.10 - Table 5.12 show the change in algorithm performance if the
galleries content is changed. In this experiment, a second standard dataset is constructed
and used. This dataset consists of five galleries of approximately 200 individuals, where
an individual is only in one gallery. Tables Table 5.10 - 5.13, summarize the results where
the algorithms are ordered by top rank score and average of rank scores, respectively. For
example, as indicated in Table 5.10, the QPCA algorithm scored highest on gallery of (Set
2). While Table 5.10 -Table 5.11 report results for the FB probes whereas Table 5.12 -
Table 5.13 report results for the duplicate I probes. The last column in these tables
represents the overall rank or order of each algorithm which indicates the general
performance indicator.
Algorithm Ranking by Top Match Gallery Size / Scored Probes
200 /200 200 /200 200 /200 200 /200 193 /193 Overall rank
Algorithm Set 1 Set 2 Set 3 Set 4 Set 5 PCA 1 4 3 3 3 2 QPCA 1 1 2 1 1 1 Red PCA 6 3 1 2 2 2 Green PCA 5 2 5 5 2 3 Blue PCA 3 4 6 6 5 4 Avg RGB PCA 4 3 4 4 4 3 SCFT CPCA 2 5 7 7 6 5 Avg Score 0.3707 0.5843 0.4988 0.5745 0.4996
Table 5.10: Variation in identification performance using five different galleries from the FB probes (top match).
73
Algorithm Ranking by Average of rank of 50 scores Gallery Size / Scored Probes
200 /200 200 /200 200 /200 200 /200 193 /193 Overall rank
Algorithm Set 1 Set 2 Set 3 Set 4 Set 5 PCA 2 4 4 3 3 3 QPCA 1 3 3 2 2 1 Red PCA 7 2 1 1 1 2 Green PCA 5 6 6 5 5 6 Blue PCA 4 7 7 6 6 7 Avg RGB PCA 6 5 5 4 4 5 SCFT CPCA 3 1 2 7 7 4 Avg Score 0.6553 0.8434 0.8739 0.8751 0.8841
Table 5.11: Variation in identification performance using five different galleries from the FB probes (average of rank scores).
Figure 5.41: Identification performance using the FB probes of (Set 1) and rank of 50
74
Figure 5.42: Identification performance using the FB probes of (Set 2) and rank of 50
Figure 5.43: Identification performance using the FB probes of (Set 3) and rank of 50
75
Figure 5.44: Identification performance using the FB probes of (Set 4) and rank of 50
Figure 5.45: Identification performance using the FB probes of (Set 5) and rank of 50
76
Algorithm Ranking by Top Match Gallery Size / Scored Probes
200 /200 200 /200 200 /200 200 /200 193 /193 Overall rank
Algorithm Set 1 Set 2 Set 3 Set 4 Set 5 PCA 2 3 2 3 3 2 QPCA 4 3 2 1 3 2 Red PCA 5 1 1 4 1 1 Green PCA 1 4 4 1 3 2 Blue PCA 6 4 4 3 3 3 Avg RGB PCA 3 2 3 2 2 1 SCFT CPCA 7 5 5 5 4 4 Avg Score 0.0269 0.2320 0.2829 0.0535 0.2283
Table 5.12: Variation in identification performance using five different galleries from the duplicate I probes (top match).
Algorithm Ranking by Average of rank of 50 scores Gallery Size / Scored Probes
200 /200 200 /200 200 /200 200 /200 193 /193 Overall rank
Algorithm Set 1 Set 2 Set 3 Set 4 Set 5 PCA 1 5 2 2 2 2 QPCA 4 6 1 1 3 3 Red PCA 2 1 3 1 1 1 Green PCA 3 7 5 3 4 5 Blue PCA 6 4 6 5 6 6 Avg RGB PCA 5 3 4 4 5 4 SCFT CPCA 7 2 7 6 7 7 Avg Score 0.1782 0.5068 0.5907 0.2620 0.7791
Table 5.13: Variation in identification performance using five different galleries from the duplicate I probes (average of rank scores).
77
Figure 5.46: Identification performance using the duplicate I probes of (Set 1) and rank of 50
Figure 5.47: Identification performance using the duplicate I probes of (Set 2) and rank of 50
78
Figure 5.48: Identification performance using the duplicate I probes of (Set 3) and rank of 50
Figure 5.49: Identification performance using the duplicate I probes of (Set 4) and rank of 50
79
Figure 5.50: Identification performance using the duplicate I probes of (Set 5) and rank of 50
5.3.7. GD versus ID
The decidability indices for the used PCA-based algorithms are calculated and listed in
Table 5.14.
Algorithm Decidability Index (d) PCA 0.1422
QPCA 0.1470 Red PCA 0.1694
Green PCA 0.1390 Blue PCA 0.1486
Avg RGB PCA 0.1475 SCFT CPCA 0.0251
Table 5.14: Decidability index (d) for each algorithm.
Figure 5.51 shows the genuine and imposter distributions using PCA. Blue and red dashed
curves represent genuine and imposter distributions; respectively, vertical lines illustrate
80
the means of the two distributions, and the region underlie the two curves are overlapping
indicates the errors resulting from their intersection. The developed QPCA algorithm
yields genuine versus imposter distribution illustrated in Figure 5.52. It can be noted that
QPCA achieved better separation between genuine and imposter distributions with less
errors. In Figure 5.53, the distributions of using Red PCA are shown. The separation of
genuine and imposter is better than that achieved by using QPCA, and in general it has the
largest separation (i.e. decidability index). As shown in Figure 5.54, the use of Green
PCA yields less separation than PCA, with more errors. The distributions using Blue PCA
and Avg PCA are shown in Figure 5.55 and Figure 5.56, respectively. They are almost
similar, although that the separation of Blue PCA is slightly better than what is achieved
by Avg RGB PCA. In the case of SCFT CPCA, the separation is the worst separation
overall algorithms as presented by its distributions in Figure 5.57. Moreover, it can be
inferred that the separation is even far less than the closest (second lowest) algorithm of
the remaining algorithms. Definitely, this indicates the highest error rate which increased
rapidly due to the drastic overlapping between genuine an imposter distributions.
81
Figure 5.51: GD versus ID for PCA algorithm
Figure 5.52: GD versus ID for QPCA algorithm
82
Figure 5.53: GD versus ID for Red PCA algorithm
Figure 5.54: GD versus ID for Green PCA algorithm
83
Figure 5.55: GD versus ID for Blue PCA algorithm
Figure 5.56: GD versus ID for Avg RGB PCA algorithm
84
Figure 5.57: GD versus ID for SCFT CPCA algorithm
5.3.8. FAR versus FRR errors
Table 5.15 gives the EER value for each PCA-based algorithm.
Algorithm EER PCA 0.151
QPCA 0.152 Red PCA 0.130
Green PCA 0.168 Blue PCA 0.177
Avg RGB PCA 0.145 SCFT CPCA 0.457
Table 5.15: The EER of each algorithm.
Figure 5.58 shows the FAR and FRR curves when using PCA algorithm. It can be noted
that EER is small, and consequently errors are low. Where the blue curve represents the
FAR, whereas the other red dashed curve represents the FRR, and the cross point between
85
the two curves denotes the equal error rate (EER) of FAR and FRR. For the developed
QPCA algorithm, Figure 5.59 illustrates the curves. The EER value is very similar to the
one achieved with PCA, and thus QPCA, most likely, generates a close number of errors.
This observation confirms the results obtained previously from the genuine and imposter
distributions. Figure 5.60 shows the curves using Red PCA algorithm. Essentially, the
EER value is the smallest among the others algorithms. The curves when using Green
PCA and Blue PCA are illustrated in Figure 5.61 and Figure 5.62, respectively. It is clear
that, EER of both ore large compared to the previous mentioned algorithms (PCA, QPCA
and Red PCA). Consequently, there are both have more errors. While Green PCA has
slightly less error than Blue PCA. The curves using Avg RGB PCA are shown in Figure
5.63. The value of EER is located between Red PCA and Green PCA, where it is greater
than the first one and smaller than the other. Finally, the curves of SCFT CPCA are shown
in Figure 5.64. The far greatest EER value is obtained, which indicates that the highest
error rate is achieved by this algorithm.
86
Figure 5.58: FAR versus FRR curve for PCA algorithm
Figure 5.59: FAR versus FRR curves for QPCA algorithm
87
Figure 5.60: FAR versus FRR curves for Red PCA algorithm
Figure 5.61: FAR versus FRR curves for Green PCA algorithm
88
Figure 5.62: FAR versus FRR curves for Blue PCA algorithm
Figure 5.63: FAR versus FRR curves for Avg RGB PCA algorithm
89
Figure 5.64: FAR versus FRR curves for SCFT CPCA algorithm
5.3.9. ROC Curves
Figure 5.65 shows the ROC curves when using PCA and QPCA algorithm. Where the
relationship between FAR and FRR can be observed clearly As FAR is increased, FRR is
decreased and accordingly (1- FRR) is increased, and vice versa. Where the small black
circle denotes the estimated EER point where the FAR and FRR are equivalent. The
obtained QPCA curve, Compared to the PCA curve, indicate that QPCA achieves higher
accuracy than PCA with accepting the same number of errors. In Figure 5.66, the curve
using Red PCA algorithm. It can be observed that the curve is the highest curve overall
others, and thus it yields the least errors. This confirms the finding from genuine and
imposter distributions and FAR and FRR curves. Figure 5.67 and Figure 5.68 show the
90
result when using Green PCA algorithm and Blue PCA Algorithm, respectively. It is clear
that, they are both lower than PCA and QPCA curves, which means more errors are
achieved. While Green PCA is bitter that Blue PCA with less error. The curve using Avg
RGB PCA is shown in Figure 5.69. It can be inferred from this curve that it achieves
results with fewer errors compared with Green and Blue PCA, but still lower that Red
PCA. Finally, the lowest curve overall which yields the highest number of errors is SCFT
CPCA as illustrated in the curve of Figure 5.70.
Figure 5.65: PCA and QPCA algorithms ROC curves
91
Figure 5.66: ROC curve for Red PCA algorithm
Figure 5.67: ROC curve for Green PCA algorithm
92
Figure 5.68: ROC curve for Blue PCA algorithm
Figure 5.69: ROC curve for Avg RGB PCA algorithm
93
Figure 5.70: ROC curve for SCFT CPCA algorithm
94
CHAPTER 6
6. CONCLUSION AND FUTURE WORK
This chapter presents the conclusions of the aspects studied in this thesis along with a
summary of the research and findings in light of experimental work and testing results.
Furthermore, proposed directions for the extension of this work are offered as possible
future work.
6.1. Summary and Findings
In this thesis, an overall background on biometric traits and systems, and on face
recognition concepts, techniques and processing have been provided. Holistic
quaternionic representation for color face images has been investigated. Besides, an
overview of Quaternion (Hypercomplex) concepts including basic definitions, properties,
and representations has been focused on. Then, the PCA and SVD concepts have been
explained and further extended to the QPCA and QSVD which have been adopted and
employed to design a biometric-based system for color face recognition and matching as
well. After that, the design and implementation used to evaluate and compare the
proposed technique with other PCA-based techniques and the testing methodology
including data and experiments have been described and discussed. Within the
performance evaluation, the following experimental investigation avenues have been
conducted:
1. Identification performance against probe categories.
95
2. Performance comparison using CMC in different resolutions.
3. Variation in identification performance.
4. GD versus ID.
5. FAR versus FRR errors.
6. ROC curves.
The following are the major contributions of this research work.
A novel scheme for efficient color face representation and processing for face-based
biometric recognition system.
The concept of the PCA and SVD are extended to the quaternionic and carried over
into the hypercomplex to compactly represent color face images.
Detailed analysis of the incorporation of color information in face recognition is
provided.
Performance improvement by using color information and the Q-PCA method.
Extended performance evaluation is preformed and comparisons from different
perspectives are provided.
The following work was performed in support of these contributions:
Development of a new algorithm for face recognition and matching based on
quaternion principle component analysis (QPCA) technique and extension of the face
recognition software to color images.
96
Implementation of other standard PCA-based algorithms for the sake of comparison
against the Q-PCA algorithm, these algorithms are the PCA, Red PCA, Green PCA,
Blue, Avg RGB PCA, and SCFT PCA.
Organization of facial data galleries and probes under different categories to be using
in the conducted experiments.
Performance testing and analysis by performing a number of stress test experiments
for these different compared implementations to evaluate and compare each with the
others.
Generation of prober evaluation graphs and tables demonstrating and representing
experimental outcomes.
The following are the major findings derived through this research work:
Use of color information can increase accuracy.
Regarding color face images, the red-channel or component is the highest effective
channel of the three color channels as concluded form its high performance, the best
separation overall algorithms, and the least EER over others. Hence, red is the most
correlated component and the most proper one to build PCA features.
While making use of the high-effective single color channel enhances the performance
of face recognition, further making use of the entire color channels holistically and
cooperatively is improving the performance more.
97
Extending the standard grayscale PCA features to the Q-PCA using quaternion color
representation can increase accuracy.
The Q-PCA has good separation between genuine and imposter distributions and
fewer errors than the PCA, despite of that Q-PCA has EER very similar to the PCA.
The experimental outcomes indicate in most cases that the proposed Q-PCA achieves
better performance than the other typical tested PCA-based algorithms.
From robustness and reliability point of views, where each algorithm has been tested
over special sets of facial database consists of samples represent the difficulties cases
such as impose and illumination. Overall the tested algorithms the Q-PCA is to
provide better robustness against such unconstrained cases.
Using face image data with higher resolution can increase the precision of the derived
PCA features and consequently can increase the identification and matching accuracy.
6.2. Future Work
Since the lack of research using Quaternion numbers to represent and process color
images. Because it is a complicated task, it has not received enough attention in the
research literature. Here are some of the suggested future studies to enhance the present
research:
The performance and accuracy can be improved by performing additional
preprocessing operations for the purpose of preparing more normalized data images,
98
which can reduce the variety of face images (such variety extremely exists in current
used database). Thus some preprocessing can be performed properly for example: all
images can be translated, rotated, and scaled so that the center of the eyes can be
placed on specific pixels, moreover faces can be masked to remove background and
hair and so on. Thereby accuracy is expected to be increased.
The current developed Q-PCA algorithm for color face recognition can be adapted,
modified and applied on another biometric trait such as color iris.
The proposed face-based biometric system can be applied using the quaternionic
representation for color images in other color spaces such as XYZ, YUV or YCbCr
color spaces. For example a face image in YCbCr color space can be represented as:
0 . . . (6.1)
The proposed Q-PCA algorithm is fully automatic. Therefore it is suggested that to
investigate the development of partially automatic QPCA-based algorithm for face
recognition. Such partially automatic algorithm is supplied with some information as
guidance such as the coordinates of the center of the eyes (this guide is already
available in FERET database). These guides used to increase the precession then
consequently the performance and accuracy are improved. Consequently, this new
implementation approach can be compared with the current fully automatic approach.
99
REFERENCES
[1] C. Jones, "Color Face Recognition using Quaternionic Gabor Filters," PhD Dissertation,
Virginia Institute of Technology, 2004.
[2] C. Wang, B. Yin, X. Bai, and Y. Sun, "Color Face Recognition Based on 2DPCA," 19th
International Conference on Pattern Recognition (ICPR 2008), pp. 1-4, 8-11 Dec., 2008.
[3] C. Wang, and X. Bai, "Color Face Recognition Based on Revised NMF Algorithm," Second
International Conference on Future Information Technology and Management Engineering
(FITME '09), pp. 455-458, 13-14 Dec., 2009.
[4] F. Al-Qunaieer, "Color Iris Recognition and Matching using Quaternion Gabor Wavelets," MS
Thesis, KFUPM University, 2009.
[5] Face Recognition Homepage, http://face-rec.org.
[6] J. Choi, S. Yang, Y. Ro, and K. Plataniotis, "Color Effect on the Face Recognition with Spatial
Resolution Constraints," 10th IEEE International Symposium on Multimedia (ISM 2008), pp.
294-301, 15-17 Dec., 2008.
[7] J. Choi, Y. Ro, and K. Plataniotis, "Color Face Recognition for Degraded Face Images," IEEE
Transactions on Systems, Man, and Cybernetics, Part B: Cybernetics, vol. 39, no. 5, pp.
1217-1230, Oct., 2009.
[8] J. Yang, and C. Liu, "A General Discriminant Model for Color Face Recognition," IEEE 11th
International Conference on Computer Vision (ICCV 2007), pp. 1-6, 14-21 Oct., 2007.
[9] C. Jones, and A. Abbott, "Color Face Recognition by Hypercomplex Gabor Analysis," 7th
International Conference on Automatic Face and Gesture Recognition (FGR 2006), pp.6-131,
2-6 April., 2006.
100
[10] N. Le Bihan, and S. Sangwine, "Quaternion Principal Component Analysis of Color Images,"
International Conference on Image Processing Proceedings (ICIP 2003), vol.1, pp. I- 809-12,
14-17 Sept., 2003.
[11] L. Ding, and H. Feng, "Quaternion K-L Transform and Biomimetic Pattern Recognition
Approaches for Color-face Recognition," IEEE International Conference on Intelligent
Computing and Intelligent Systems (ICIS 2009), vol.1, pp. 165-169, 20-22 Nov., 2009.
[12] H. Moon, and P. Phillips, “Computational and Performance Aspects of PCA-based Face
Recognition Algorithms,” Perception, vol. 30, pp. 301-321, 2001.
[13] S. Pei, and C. Cheng, “A Novel Block Truncation Coding of Color Images by Using
Quaternion-moment Preserving Principle,” IEEE International Symposium on Circuits and
systems, vol. 2, pp. 684-687, Atlanta, FA. 12-15 May., 1996.
[14] S. Li, and A. Jain, Handbook of Face Recognition, Springer, New York, USA, 2005.
[15] S. Chen, Y. Sun, and B. Yin, "A Novel Hybrid Approach Based on Sub-pattern Technique
and Extended 2DPCA for Color Face Recognition," 11th IEEE International Symposium on
Multimedia (ISM '09), pp. 630-634, 14-16 Dec., 2009.
[16] P. Shih, and C. Liu, "Extracting Efficient Color Features for Face Recognition Using
Evolutionary Computation," 6th International Conference on Computational Intelligence and
Multimedia Applications, pp. 285- 290, 16-18 Aug., 2005.
[17] J. Shlens, “A Tutorial on Principal Component Analysis,” Unpublished, version 3.01, 2009.
[18] L. Smith, “A Tutorial on Principal Component Analysis,” Unpublished, 2002.
[19] T. Ganapathi, and K. Plataniotis, “Color Face Recognition under Various Learning
Scenarios,” IEEE Canadian Conference on Electrical and Computer Engineering, 5-7 May.,
2008.
101
[20] T. Ganapathi, "Color Image Based Face Recognition," MS Thesis, University of Toronto,
2009.
[21] M. Thomas, C. Kambhamettu, and S. Kumar, "Face Recognition Using a Color Subspace
LDA Approach," 20th IEEE International Conference on Tools with Artificial Intelligence
(ICTAI '08), vol. 1, pp. 231-235, 3-5 Nov., 2008.
[22] M. Turk, and A. Pentland, “Eigenfaces for recognition,” Journal of Cognitive Neuroscience,
vol. 3, no 1, pp. 72–86, 1991.
[23] L. Wei, X. Yi, X. Yang, and L. Song, "Local Quaternionic Gabor Binary Patterns for Color
Face Recognition," IEEE International Conference on Acoustics, Speech and Signal
Processing (ICASSP 2008), pp. 741-744, March 31 2008-April 4 2008.
[24] W. Zhao, R. Chellappa, J. Phillips, and A. Rosenfeld, “Face Recognition: A Literature
Survey,” ACM Computing Surveys, vol. 35, no. 4, pp. 399-458, Dec., 2003.
[25] S. Thakur, J.K. Sing, D. K. Basu, M. Nasipuri, and M. Kundu, “Face Recognition Using
Principal Component Analysis and RBF Neural Networks,” First International Conference on
Emerging Trends in Engineering and Technology (ICETET '08), pp. 695-700, 2008.
[26] T. Rahman, and A. Bhuiyan, “Face Recognition using Gabor Filters ,” 11th International
Conference on Computer and Information Technology Proceedings (ICCIT 2008), pp. 510-
515, Khulna, Bangladesh, 25-27 Dec., 2008.
[27] X. Wang, C. Huang, G. Ni, and J. Liu, “Face Recognition Based on Face Gabor Image and
SVM,” 2nd International Congress on Image and Signal Processing (CISP '09), pp. 1-4, Oct.,
2009.
[28] C. Liu, and D. Dai, “Face Recognition Using Dual-Tree Complex Wavelet Features,” IEEE
Transactions On Image Processing, vol. 18, no. 11, pp. 2593-2599, Nov., 2009.
102
[29] Z. Chai, K. Ma, and Z. Liu, “Complex Wavelet-based Face Recognition Using Independent
Component Analysis,” 5th International Conference on Intelligent Information Hiding and
Multimedia Signal Processing, pp. 832-835, 12-14 Sep.,2009.
[30] L. Shi, “Exploration in Quaternion Colour,” MS Thesis, Simon Fraser University, 2005.
[31] S. Sangwine, and N. Le Bihan, “Computing the SVD of a quaternion matrix,” 7th IMA
Conference on Mathematics in Signal Processing Royaume-Uni, 2006.
[32] P. J. Phillips, H. Moon, P. Rauss, and S. Rizvi, “The FERET Evaluation Methodology for
Face-Recognition Algorithms,” Tech. Rep. NISTIR 6264, National Institute of Standards and
Technology, 7 Jan. 1999.
[33] P. J. Phillips, A. Martin, C. Wilson and M. Przybocki. “An Introduction to Evaluating
Biometric Systems”, IEEE Computer, pp. 56-63, Feb., 2000.
[34] F. Zhang, "Quaternions and Matrices of Quaternions," Linear algebra and its applications,
vol. 251, pp. 21-57, 1997.
[35] A. Koschan and M. Abidi, “Digital color image processing,” first ed. John wiley &
Sons, new york, 2008.
[36] W. Hamilton, “On quaternions; or on a new system of imaginaries in algebra,” Edinburgh
and Dublin Philosophical Magazine and Journal of Science, pp. 25–36, 1844–1850.
[37] A. McCabe, T. Caelli, G. West and A. Reeves, “Theory of spatiochromatic image encoding
and feature extraction,” Journal of the Optical Society of America A: Optics, Image Science,
and Vision, vol. 17, no. 10, 2000.
[38] S.-C. Pei, J.-J. Ding and J.-H. Chang, “Efficient implementation of quaternion Fourier
transform, convolution, and correlation by 2-D complex FFT,” IEEE Transactions on Signal
Processing, vol. 49, no. 11, 2001.
103
[39] S. Westland and C. Ripamonti, “Computational Colour Science Using MATLAB,” John
Wiley & Sons, new york, 2004.
[40] A. McCabe, T. Caelli, G. West and A. Reeves, “ Encoding and Processing Spatio-Chromatic
Image Information Using Complex Fourier Transform Methods,” Technical Report, Curtin
University of Technology, Perth, Western Australia, pp. 228–24, 1997.
104
Vita
Personal Information
Date of Birth: 1982 Place of Birth: Jeddah Nationality: Saudi Emails:
1. [email protected] 2. [email protected]
Mobile Number: 0504696694 Present Address: 1st street, Al-Badee District, Dammam,
Saudi Arabia. Permanent Address: P.O.BOX 1237, Jeddah – 21433,
Saudia Arabia. Languages:
1. Arabic (native language) 2. English
Education
2000 -2004: King Abdul-Aziz University Jeddah, Bachelor of Science in Computer Science
2007-2010: King Fahd University of Petroleum & Minerals Dhahran, Master of Science in Computer Science
Work
2004-2005: Developer in IT Unit in Deanship of Community Services & Continuing Education at King Abdul-Aziz University
2005-Present: Graduate Assistant in Computer Science Department in College of Education at King Abdul-Aziz University
Interests
Image processing and pattern recognition Biometric Systems Web development